Manual harvesting bell peppers as a greenhouse product is time-consuming and has high labor costs. Various two-dimensional machine vision systems have been developed for the robotic harvest of this product. Due to the complex environment of the greenhouse, high density of branches and leaves, the presence of pests and diseases and sunburn, non-uniform and variable light conditions, and asynchronous maturation of the crop lead to non-uniformity in color. Therefore, harvesting colored sweet peppers are also associated with challenges in addition to green sweet peppers. The aim of this study is to recognize sweet peppers in the point cloud model based on the Fast Point Features Histogram (FPFH) descriptor, and machine learning. Kinect V2 sensor was used to record depth images. After creating a 3-D model, geometric features (FPFH) were used for classification, and H color channel was used to create supervision in 9 classification models on 15 3-D color peppers models supervised. According to the F1-Score criteria, the area under the roc curve, and the accuracy of the algorithm, the K-nearest neighbor classifier model was selected as the best model. To evaluate the KNN algorithm, 15 3D models of red, orange and yellow pepper were performed and finally, they were counted manually. The results showed that the value of F1-Score was higher than 0.7 in all cases, except for one case, the area under the roc curve was higher than 0.8 in all cases and the accuracy of the classification algorithm was higher than 0.9 in all cases. In addition, the human evaluation results showed that the minimum recognition accuracy is 71.42%, and the maximum is 100%.